International Journal of Heat and Mass Transfer 100 (2016) 876–891

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

journal homepage: www.elsevier.com/locate/ijhmt

Large scale three-dimensional topology optimisation of heat sinks cooled

by natural convection
Joe Alexandersen ⇑, Ole Sigmund, Niels Aage
Department of Mechanical Engineering, Solid Mechanics, Technical University of Denmark, Nils Koppels Allé, Building 404, DK-2800 Kgs. Lyngby, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: This work presents the application of density-based topology optimisation to the design of three-
Received 18 August 2015 dimensional heat sinks cooled by natural convection. The governing equations are the steady-state
Received in revised form 2 May 2016 incompressible Navier–Stokes equations coupled to the thermal convection–diffusion equation through
Accepted 4 May 2016
the Bousinessq approximation. The fully coupled non-linear multiphysics system is solved using sta-
bilised trilinear equal-order finite elements in a parallel framework allowing for the optimisation of large
scale problems with order of 20–330 million state degrees of freedom. The flow is assumed to be laminar
Keywords:
and several optimised designs are presented for Grashof numbers between 103 and 106 . Interestingly, it is
Topology optimisation
Heat sink design
observed that the number of branches in the optimised design increases with increasing Grashof num-
Natural convection bers, which is opposite to two-dimensional topology optimised designs. Furthermore, the obtained
Large scale topologies verify prior conclusions regarding fin length/thickness ratios and Biot numbers, but also indi-
Multiphysics optimisation cate that carefully tailored and complex geometries may improve cooling behaviour considerably com-
pared to simple heat fin geometries.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction arrangement; Iyengar and Bar-Cohen [3] investigate vertical

Natural convection is the phenomena where density-gradients
due to temperature differences cause a fluid to move. Natural con-
vection is therefore a natural way to passively cool a hot object,
such as electronic components, light-emitting diode (LED) lamps
or materials in food processing.
Structural optimisation is the discipline of modifying the design
of a structure in order to improve its performance with respect to
some desirable behaviour. Structural optimisation techniques,
such as size and configuration optimisation, are very efficient if a
close-to-optimal design is already known, or the topology of the
structure is dictated by e.g. certain manufacturing methods. These
methods are frequently applied to the design of heat sinks in elec-
tronics cooling. The following literature review is by no means
complete, giving only representative examples: Morrison [1] opti-
mises plate fin heat sinks in natural convection using a downhill
simplex method and empirical correlations. The design variables
are the fin thickness, fin spacing and backplate thickness; Ladezma
and Bejan [2] investigate the geometric arrangement of staggered
vertical plates in natural convection using numerical simulations.
The design variables are various dimensions of the staggered
⇑ Corresponding author.
E-mail addresses: joealex@mek.dtu.dk (J. Alexandersen), sigmund@mek.dtu.dk
(O. Sigmund), naa@mek.dtu.dk (N. Aage).
pin-fin, plate-fin and triangular-fin heat sinks in natural convection
using analytical and empirical correlations. The design variables
are the fin thickness and spacing; Bahadur and Bar-Cohen [4]
optimise staggered pin-fin heat sinks for natural convection cooled
microprocessor applications using analytical equations. The design
variables are pin height, diameter and spacing; Jang et al. [5]
optimise radial pin-fin heat sinks for LED applications using
numerical simulation and a genetic algorithm. The design variables
are the number and length of fins. Furthermore, there exists a vast
literature treating optimal structures for surface-to-point and
volume-to-point heat generation, e.g. [6], but these do not consider
convective heat transfer and are thus not directly relevant to the
problems at hand.
While parameter studies and simple optimisation techniques,
like the abovementioned, can provide insight and improvements
to existing designs, they are all limited in their design freedom
as an a priori determined initial design must be given. Topology
optimisation allows for a vastly expanded design space, allowing
for the formation of unintuitive and unanticipated designs that
fully exploit the governing physics. Topology optimisation is a
material distribution method [7–9] used to optimise the layout
of a structure. In order to take convective heat transfer, to an
ambient fluid, into account in the design process of density-
based methods, a common extension is to introduce some form
of interpolation of the convection boundaries, see e.g. [10–13].

http://dx.doi.org/10.1016/j.ijheatmasstransfer.2016.05.013
0017-9310/Ó 2016 Elsevier Ltd. All rights reserved.
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891 877

More recently, these simplified models have been used by Dede In recent years, an increasing body of work has been published
et al. [14] to design and manufacture heat sinks subject to jet on efficient large scale topology optimisation. These works cover
impingement cooling, as well as by Zhou et al. [15] in an industrial the use of high-level scripting languages [40,41], multiscale/
framework to optimise electric motor covers and heat sinks. -resolution approaches [42,43] and parallel programming using
Generally, the application of a predetermined and design- the message parsing interface (MPI) and C/Fortran [44–47]. To
independent convection coefficient is at best inaccurate and may facilitate the solution to truly large scale conjugate heat transfer
have a strong influence on resulting designs and their performance. problems, the implementation in this article is done using PETSc
In practise, topology optimisation based on simplified models may [48] and the framework for topology optimisation presented in
lead to unanticipated designs and closed cavities, thereby violating [47].
the assumptions of the simplified model [16]. During the optimisa- The layout of the article is as follows: Section 2 presents the
tion process, the design changes significantly and the interaction governing equations; Section 3 presents the topology optimisation
with the ambient fluid changes as well. Therefore, to ensure phys- problem; Section 4 briefly discusses the finite element formula-
ically correct capturing of the aspects of convective heat transfer, tion; Section 5 discusses the numerical implementation details;
the full conjugate heat transfer problem must be solved. Obviously, Section 6 presents scalability results for the parallel framework,
the employment of a full-blown fluid model increases computa- optimised designs for two test problems, as well as verification
tional time and complexity considerably. Hence, using the simpli- results; Section 7 finishes with a discussion and conclusion.
fied convection approach, that provides for very fast solution times,
may likely provide a good first estimate for an optimised topology 2. Governing equations
or may be used for post-processing once topology and associated
local convection coefficients have been found. Furthermore, it The dimensionless form of the governing equations have been
may be beneficial to use the simplified convection approach, when derived based on the Navier–Stokes and convection–diffusion
the flow is too complex to model in current optimisation settings equations under the assumption of constant fluid properties,
(hundreds or thousands of function evaluations). Please see incompressible, steady flow and neglecting viscous dissipation.
[13,15–18] for further discussions on the strengths and weak- Furthermore, the Boussinesq approximation has been introduced
nesses of using the simplified convection approach. to take density-variations due to temperature-differences into
Topology optimisation for fluid systems began with the treat- account. A domain is decomposed into two subdomains,
ment of Stokes flow in the seminal article by Borrvall and Petersson X ¼ Xf [ Xs , where Xf is the fluid domain and Xs is the solid
[19] and has since been applied to Navier–Stokes [20], as well as domain. In order to facilitate the topology optimisation of conju-
passive transport problems [21,22], reactive flows [23], transient gate natural convective heat transfer between a solid and a sur-
flows [24–26], fluid–structure interaction [27,28], amongst many rounding fluid, the equations are posed in the unified domain, X,
others. The extension of topology optimisation to turbulent fluid and the subdomain behaviour is achieved through the control of
flow is very much in its infancy [29] and requires further research. coefficients. The following dimensionless composite equations
Conjugate heat transfer problems were first treated in [30,31] and are the result.
is very much an active field of research today [32–37]. However, 8x 2 X:
almost all work is focused on forced convection, where the fluid

@ui @ @ui @uj @p
flow is induced by a fan, pump or pressure-gradient. The authors uj  Pr þ þ ¼ aðxÞui  GrPr 2 eig T ð1aÞ
have previously presented a density-based topology optimisation @xj @xj @xj @xi @xi
approach for two-dimensional natural convection problems [18]. @uj
¼0 ð1bÞ
Recently, Coffin and Maute presented a level-set method for @xj


steady-state and transient natural convection problems using the @T @ @T
eXtended finite element method (X-FEM) [38]. Interested readers uj  KðxÞ ¼ sðxÞ ð1cÞ
@xj @xj @xj
are referred to [18] for further references and a deeper introduc-
tion to topology optimisation in fluid dynamics and heat transfer. where ui is the velocity field, p is the pressure field, T is the temper-
Throughout this article, the flows are assumed to be steady and ature field, xi denotes the spatial coordinates, eig is the unit vector in
laminar. The fluid is assumed to be incompressible, but buoyancy the gravitational direction, aðxÞ is the spatially-varying effective
effects are taken into account through the Boussinesq approxima- impermeability, KðxÞ is the spatially-varying effective thermal con-
tion, which introduces variations in the fluid density due to tem- ductivity, sðxÞ is the spatially-varying volumetric heat source term,
perature gradients. The inclusion of a Brinkman friction term Pr is the Prandtl number, and Gr is the Grashof number.
facilitates the topology optimisation of the fluid flow. The effective thermal conductivity, KðxÞ, is defined as:
The scope of this article is primarily to present and provide (
1 if x 2 Xf
basic verifications of a large scale, three-dimensional framework KðxÞ ¼ ð2Þ
for topology optimisation of thermal heat sinks. The methodology
1
Ck
if x 2 Xs
builds on the two-dimensional framework presented in [18]. Thus, k
only a brief overview of the underlying finite element and topology where C k ¼ kfs is the ratio between the fluid thermal conductivity, kf ,
optimisation formulation is given and the reader is referred to [18] and the solid thermal conductivity, ks . Theoretically, the effective
for further information. The numerical extension to three dimen- impermeability, aðxÞ, is defined as:
sions is non-trivial and hence the present article includes new dis- 
0 if x 2 Xf
cussions on interpolation and continuation strategies, as well as on aðxÞ ¼ ð3Þ
1 if x 2 Xs
computational issues arising from solving the large-scale, non-
linear equation systems considered. In this first application of in order to ensure zero velocities inside the solid domain. However,
topology optimisation to natural convection problems in three numerically this requirement must be relaxed as will be described
dimensions, dimensionless parameters and fictitious properties in Section 3. The volumetric heat source term is defined as being
are used. Nevertheless, interesting insight will be gained on the active within a predefined subdomain of the solid domain, x  Xs :
effect of the Grashof number in optimal design. Ongoing and future 
0 if x R x
work is devoted to the treatment of physically realistic problems sðxÞ ¼ ð4Þ
for LED lamp coolers [39] and other practical devices. s0 if x 2 x
878 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891

2
where s0 ¼ kqLDT is the dimensionless volumetric heat generation, q is
f
the dimensional volumetric heat generation, DT is the reference
temperature difference and L is the reference length scale.
The Prandtl number is defined as:
m 3.3. Continuation scheme
Pr ¼
C
where m is the kinematic viscosity, or momentum diffusivity, and C
is the thermal diffusivity. It thus describes the relative spreading of
viscous and thermal effects. The Grashof number is defined as:
gb DT L3
Gr ¼
m2
where g is the acceleration due to gravity and b is the volumetric
coefficient of thermal expansion. It describes the ratio between
the buoyancy and viscous forces in the fluid. The Grashof number
is therefore used to describe to what extent the flow is dominated
by natural convection or diffusion. For low Gr the flow is dominated
by viscous diffusion and for high Gr the flow is dominated by natu-
ral convection. The problems in this article are assumed to have
large enough buoyancy present to exhibit natural convective
effects, but small enough Gr numbers to exhibit laminar fluid
motion.
3. Optimisation formulation
3.1. Interpolation functions
In order to perform topology optimisation, a continuous design
field, cðxÞ, varying between 0 and 1 is introduced. Pure fluid is rep-
resented by cðxÞ ¼ 1 and solid by cðxÞ ¼ 0. For intermediate values
between 0 and 1, the effective conductivity is interpolated as
follows:
cðC k ð1 þ qf Þ  1Þ þ 1
KðcÞ ¼
C k ð1 þ qf cÞ
and likewise the effective impermeability is interpolated using:
1c
aðcÞ ¼ a
1 þ qa c
The interpolation functions ensure that the end points defined in (4)
and (5), respectively, are satisfied. The effective impermeability is
bounded to a in the solid regions and this upper bound should be
chosen large enough to provide vanishing velocities, but small
enough to ensure numerical stability. The convexity factors, qf
and qa , are used to control the material properties for intermediate
design values in order to promote well-defined designs without
intermediate design field values.
The design field is regularised using a PDE-based (partial differ-
ential equation) density filter [49,47] and the optimisation prob-
lem is solved using the method of moving asymptotes (MMA)
[50,47].
ð5Þ
A continuation scheme is performed on various parameters in
order to stabilise the optimisation process and to improve the opti-
misation results. It is the experience of the authors that the pro-
vided continuation scheme yields better results than starting
with the end values, although this cannot generally be proven
ð6Þ
[51,52].
The chosen continuation strategy consists of five steps:
qf 2 f0:881; 8:81; 88:1; 881; 881g ð10aÞ
qa 2 f8; 8; 8; 98; 998g ð10bÞ
n o
a 2 105 ; 105 ; 105 ; 106 ; 107 ð10cÞ
The sequence is chosen in order to alleviate premature convergence
to poor local minima. The value of qf is slowly increased to penalise
intermediate design field values with respect to conductivity. The
maximum effective permeability, a, is set relatively low during
the first three steps, as this ensures better scaled sensitivities and
more stable behaviour. Over the last two steps, a is increased by
two orders of magnitude in order to further decrease the velocity
magnitudes in the solid regions. The particular values of qf and qa
are chosen by empirical inspection, such as to ensure the approxi-
mate collocation of the fluid and thermal boundaries. The suggested
continuation strategy is based on extensive numerical studies and
experiments reflecting that: (a) a high initial a causes designs to
glue to outer walls of the design domain; (b) a high initial qf causes
too rapid convergence to 0–1 solutions; (c) starting with final val-
ues for all parameters gives bad scaling of initial sensitivities and
ð7Þ hence convergence towards inferior local optima. Here, it is impor-
tant to note that the optimisation problem is by no means convex
and any optimised design will at best be a local minimum. The
obtained design will always depend on the initial design, as well
ð8Þ as the continuation strategy. However, in the authors experience,
the chosen continuation strategy gives a good balance between con-
vergence speed, final design performance and physicality of the
modelling. The effects of the steps of the current continuation strat-
egy on the design distribution will be discussed in Section 6.5. Also,
it is noted that results obtained using the suggested continuation
strategy: (a) indicate that consistent and well-performing designs
are obtained as verified by cross-checks; (b) indicate that boundary
effects are captured sufficiently accurately as verified by COMSOL
runs based on body-fitted meshes and accurate boundary condi-
tions, see Section 6.4.

3.2. Optimisation problem

4. Finite element formulation
The topology optimisation problem is defined as:
Z The governing equations are discretised using stabilised trilin-
minimise : f ðc; T Þ ¼ sðxÞ T dV ear hexahedral finite elements. The design field is discretised using
c2D elementwise constant variables, in turn rendering the effective
Z x Z
thermal conductivity and the effective impermeability to be ele-
subject to : g ðcÞ ¼ 1  c dV 6 v f dV ð9Þ
Xd Xd mentwise constant. The monolithic finite element discretisation
Rðc; u; p; T Þ ¼ 0 of the problem ensures continuity of the temperature field, as well
as the fluxes across fluid–solid interfaces. The particularities of the
0 6 cðxÞ 6 1 8x 2 Xd
implemented finite element formulation are as detailed in [18].
where c is the design variable field, D is the design space, f is the However, simpler stabilisation parameters have been used in order
thermal compliance objective functional, g is the volume constraint to allow for consistent sensitivities to be employed, see Appendix
functional, Rðc; u; p; T Þ is the residual arising from the stabilised A. The Jacobian matrix is now fully consistent, in that variations
weak form of the governing equations, and Xd # X is the design of the stabilisation parameters with respect to design and state
domain. fields are included, in contrast to the original work [18].
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891
5. Numerical implementation
The discretised FEM equations are implemented in PETSc [48]
based on the topology optimisation framework presented in [47].
The PETSc framework is used due to its parallel scalability, the
availability of both linear and non-linear solvers, preconditioners
and structured mesh handling possibilities. All components
described in the following are readily available within the PETSc
framework.
5.1. Solving the non-linear system
The non-linear system of equations is solved using a damped
Newton method. The damping coefficient is updated using a poly-
nomial L2 -norm fit, where the coefficient is chosen as the min-
imiser of the polynomial fit. The polynomial fit is built using the
L2 -norm of the residual vector at the current solution point, at
50% of the Newton step and at 100% of the Newton step. This
residual-based update scheme combined with a good initial vector
(the solution from the previous design iteration) has been observed
to be very robust throughout the optimisation process for the mod-
erately non-linear problems treated. To further increase the
robustness of the non-linear solver, if the Newton solver fails from
the supplied initial vector, a ramping scheme for the heat genera-
tion magnitude is applied in order to recover. Throughout, the con-
vergence criteria for the Newton solver is set to a reduction in the
2
L -norm of the residual of 10 4
relative to the initial residual.
5.2. Solving the linear systems
Due to the large scale and three-dimensional nature of the trea-
ted problems, the arising linearised systems of equations and their
solutions are by far the most time consuming part of the Newton
scheme. Therefore, to make large scale problems tractable the
(unsymmetric) linear systems are solved using a fully parallelised
iterative Krylov subspace solver.
Constructing an iterative solver that is independent of problem
settings and at the same time possesses both parallel and numer-
ical scalability, is intricate and beyond the scope of this work,
where focus is on the optimisation. However, in order to facilitate
the solution of large scale optimisation problems, an efficient sol-
ver is required. To this end, the authors use the readily available
core components in PETSc to construct a solver with focus on sim-
plicity, as well as reduced wall clock time. This is quite easy to
obtain for solvers and preconditioners that rely heavily on
matrix–vector multiplications. To this purpose the flexible GMRES
(F-GMRES) method is used as the linear solver combined with a
geometry-based Galerkin-projection multigrid (GMG) precondi-
tioner. Such a solver depends highly on the quality of the smoother
to guarantee fast convergence [53]. The authors have found that a
simple Jacobi-preconditioned GMRES provides a reasonable choice
of smoother.1 The convergence criterion for the Krylov solver is set
to 105 relative to the initial residual.
6. Results
6.1. Problem setup
The considered optimisation problem is an (academic) example
of a heat sink in a closed cubic cavity. Fig. 1 shows the problem
setup with dimensions and boundary conditions. The heat source
(black in figure), exemplifying an electronics chip, is placed in
1
Due to the choice of a Krylov smoother, the multigrid preconditioner will vary
slightly with input and thus require a flexible outer Krylov method.
879
(a) Dimensions (b) Boundary conditions
Fig. 1. Illustration of the problem setup for the heat sink in a close cubic cavity. The
heat source is black and the design domain is grey.
the mid-bottom of the cavity and modelled using a small block of
solid material with volumetric heat generation, s0 ¼ 104 . The
design domain (grey in figure) is placed on top of the heat source
in order to allow the cooling fluid to pass underneath it, as well
as to allow room for wires etc. The vertical and top outer walls
of the cavity are assumed to be kept at a constant cold tempera-
ture, T ¼ 0, while the bottom is insulated. The height of the cavity
has been used as the reference length scale, L. Thus, the cavity
dimensions are 1  1  1, the design domain dimensions are
0:75  0:75  0:75 and the heat source dimensions are
0:1  0:05  0:1. A discussion on the definition of the temperature
difference in the Grashof number can be found in Appendix B.
Initial investigations showed that due to the symmetry of the
domain and boundary conditions, the design and state solutions
remained quarter symmetric throughout the optimisation. Thus,
the computational domain is limited to a quarter of the domain
with symmetry boundary conditions. The volume fraction is set
to 5%, i.e. v f ¼ 0:05, for all examples.
6.2. Parallel performance
To demonstrate the parallel performance of the state solver, the
optimisation problem is solved on a fixed mesh for different num-
bers of processes. All computations in this article were performed
on a cluster, where each node is equipped with two Intel Xeon e5-
2680v2 10-core 2.8 GHz processors. The results shown in Tables 1
and 2 are averaged over 250 design iterations and show the perfor-
mance for Gr ¼ 103 and Gr ¼ 106 , respectively. The data shows that
the proposed solver scales almost linearly in terms of speed up, and
more importantly that the performance is only slightly affected by
the Grashof number.
In order to quantify the degree of numerical scalability, a sec-
ond study is performed in which the mesh resolution is varied.
The study is conducted for both low and high Grashof numbers
and the average F-GMRES iterations are collected in Table 3. The
total number of design iterations averaged over was 250, 500
and 1000, respectively, for the three mesh resolutions. The data
clearly shows that the computational complexity increases with
problem size, and thus that the solver is not numerically scalable.
However, since the growth in numerical effort, i.e. the number of
880 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891

Table 1 as will be verified in Section 6.4. The general characteristics of all

Average time taken per state solve over 250 design iterations for Gr ¼ 103 at a mesh the designs are similar, i.e. all designs are ‘‘thermal trees” with con-
resolution of 80  160  80.
ductive branches moving the heat away from the heat source.
Processes Time [s] Scaling However, it can clearly be seen that the design changes consider-
160 53.2 1.00 ably with increasing convection-dominance (increasing Gr). For
320 28.9 0.54 increasing Gr the conductive branches contract, resulting in a smal-
640 14.1 0.26 ler spatial extent of the overall heat sink. This intuitively makes
sense as the problem goes from one of conduction/diffusion at
Gr ¼ 103 to convection at Gr ¼ 106 . When diffusion dominates,
Table 2
the goal for the branches essentially becomes to conduct the heat
Average time taken per state solve over 250 design iterations for Gr ¼ 106 at a mesh
resolution of 80  160  80. directly towards the cold walls. As convection begins to matter, the
fluid movement aids the transfer of heat away from the heat sink
Processes Time [s] Scaling
and the branches do not need to be as long. Instead, the design
160 62.6 1.00 forms higher vertical interfaces in order to increase surface area
320 31.9 0.51
parallel to the flow direction and thus increase fluid velocity and
640 16.5 0.26
enhance overall heat transfer. At the same time, the complexity
of the designs increases as the importance of convection increases.
This can be seen by studying the number of primary and secondary
Table 3
Average iterations for the linear solver per state solve over entire design process for branches. Primary branches are defined as those connected to the
Gr ¼ 103 and Gr ¼ 106 at varying mesh resolutions. heat source directly and secondary branches are those connected
to primary branches. Table 4 shows the number of primary and
Mesh size Gr ¼ 103 Gr ¼ 106
secondary branches of the optimised designs shown in Fig. 2. The
80  160  80 7.5 5.6 number of primary branches is largest for the diffusion-
160  320  160 10.1 7.7
dominated case, but more or less constant thereafter. However, it
320  640  320 18.4 15.6
can be seen that the number of secondary branches significantly
increases as the Gr-number is increased. Table 4 also shows that
the total surface area3 is decreasing for the three lower
F-GMRES iterations, is moderate, it is concluded that the proposed Gr-numbers and then increases slightly.
solver is indeed applicable for solving large scale natural convec- It is interesting to note, that the trend of increasing complexity
tion problems. with increasing Gr-number is the reverse of what was observed for
two-dimensional problems [18]. There, the complexity of the
design decreased as the Gr-number increased. This difference is
6.3. Varying Grashof number due to the fact that going into three dimensions allows the fluid
to move around and through the design making it more a question
The problem is investigated for varying Gr under constant volu- of ‘‘topology”, in contrast to the two-dimensional case, where it is a
metric heat generation, s0 ¼ 104 , Prandtl number, Pr ¼ 1, and ther- question of surface shape. Physically, in two dimensions additional
mal conductivity ratio, C k ¼ 102 . The purpose of the present study branches disturb the flow and thus the heat transfer; in three
is not to provide a detailed physical example. It is rather to provide dimensions, vertical branches improve the heat transfer through
phenomenological insight into the effect of changing the governing an increased vertical surface area. Fig. 3 shows the optimised
parameter of the fluid-thermal coupling, namely the Grashof num- designs for Gr ¼ 103 and Gr ¼ 106 from below. The radial extent
ber. However, physical interpretations of tuning the Gr-number, of the designs are emphasised from these views. It is also seen that
while keeping the dimensionless volumetric heat generation and the branches for Gr ¼ 106 are positioned to ensure that the struc-
the Pr-number constant, could be e.g. equivalent to tuning the ture is open from below, with the branches forming vertical walls
gravitational strength (going from microgravity towards full grav- as discussed above.
ity) or the dimensional volumetric heat generation. It is important Table 5 shows a cross-check of the objective functions for the
to note that when interpreting the results, the dimensional tem- optimised designs. It can be seen that all designs optimised for cer-
perature scale will differ for the two interpretations. While tuning tain flow conditions outperform the other designs for the specified
the gravitational strength, the temperature scale remains the Gr-number.
same; by varying the magnitude of the volumetric heat generation, The optimisations were run for 500 design iterations for each
the temperature scale varies accordingly.2 optimised design. Table 6 shows the computational time, average
The computational mesh is 160  320  160 elements yielding non-linear iterations and linear iterations for the optimised
a total of 8; 192; 000 elements and 41; 603; 205 degrees of freedom designs in Fig. 2 using 1280 cores. It can be seen that the compu-
for the quarter domain. The design domain consists of 3; 456; 000 tational time only weakly increases as the Gr-number is increased
elements and the filter radius is set to 2:5 times the element size. and remains close to 1.2 minutes per design iteration on average. It
Fig. 2 shows the optimised designs for varying Gr-number with is interesting to note, that Gr ¼ 106 seems easier to solve than the
superimposed slices of the corresponding temperature fields. Due others as it exhibits lower average number of linear iterations than
to the use of density filtering, the interface between solid and fluid all others, as well as a lower maximum number of non-linear iter-
regions for the optimised designs are not exactly described but
ations than Gr ¼ 105 . Furthermore, it can be seen that due to the
consists of a layer of intermediate design field values. The opti-
Newton method starting from a good initial vector (solution from
mised design distributions are shown thresholded at c ¼ 0:05,
previous design iteration), only 2 non-linear iterations are needed
which is the approximate location of the computational interface
for most of the design iterations independent of Gr-number.
2
2
The dimensionless volumetric heat generation is given by s0 ¼ kqLDT . Requiring that
f
s0 remains constant, means that the dimensional volumetric heat generation, q, and
the reference temperature difference, DT, must vary in unison, i.e. DqT ¼ const:. An 3
The surface area is computed using an isosurface at the selected threshold design
increase in the Gr-number is thus achieved through the increase of DT. field value.
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891 881

Fig. 2. Optimised designs for varying Gr-number at a mesh resolution of 160  320  160.

Table 4
penalisation used to enforce boundary conditions in the topology
The number of primary and secondary branches, as well as the surface area for the
optimised designs of Fig. 2.
optimisation must be verified with an exact boundary representa-
tion; partly, the threshold value of c ¼ 0:05 for converting the
Gr Primary Secondary Surface area
continuous density map to a structure with sharp edges must be
10 3 12 0 0.887 verified; and partly, the finite element model itself must be verified
104 8 16 0.853 by a commercial or publicly available software. For these reasons,
105 8 28 0.834 this section studies the four designs optimised so far using the
106 8 48 0.846 COMSOL Multiphysics 5.2 finite element-based simulation
software.
First, the smoothed designs (isovolumes thresholded at
6.4. Verification of boundary interpolation model c ¼ 0:05) are imported into COMSOL and verifications are run with
the same parameters and Grashof numbers as for the topology
For several reasons, it is important to verify the obtained optimisation model. The full domain is analysed and for all Grashof
topology optimisation results using an alternative model. Partly, numbers, the geometry and fluid domain are meshed with a total
the voxel grid used in the topology optimisation must be verified of approximately 760; 000 tetrahedral elements, where the
by a smooth, body-fitted mesh; partly, the Brinkman-type elements are graded away from surfaces. Fig. 4 shows the
882 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891

Fig. 3. Temperature distribution in optimised designs for Gr ¼ 103 and Gr ¼ 106 at a mesh resolution of 160  320  160 – view from below.

Table 5
Cross-check objective function values for the designs shown in Fig. 2. The compliance Table 7 shows a cross-check of the objective functions obtained
is shown for the full domain, which is 4 times the values for the quarter domain. Bold using COMSOL. It can be seen that all designs optimised for certain
figures highlight the minimum objective function for a given analysis Grashof flow conditions again outperform the other designs for the speci-
number.
fied Gr-number and thus, the conclusions of Section 6.3 are valid.
Analysis Gr Optimisation Gr The objective values for the verification calculations are consis-
103 104 105 106
tently around 20% lower than the optimisation values from Table 5,
the reason for which will be discussed below.
103 8.26 8.27 8.96 9.45
In order to verify the accuracy of the topology optimisation
104 7.73 7.52 7.98 8.45
approach in further detail, the heat sink geometry for Gr ¼ 106 is
105 5.95 5.80 5.62 5.76
4.54 4.49 4.25 4.10 investigated using a finer mesh of 1; 566; 642 tetrahedral elements,
106
where refinement has mainly taken place in the fluid domain. For
this refined mesh, the objective value is 3.36, which may be com-
Table 6 pared to the value from the coarser mesh in Table 5 of 3.34. This
Computational time, average non-linear iterations and linear iterations for the small change indicates that the integral objective value indeed
optimised designs in Fig. 2. has converged, nevertheless, the finer mesh is used to better cap-
Gr Time Non-linear: avg. (max) Linear: contin. steps – avg. (max) ture local boundary effects in the following discussion. Fig. 5 shows
3 9:56 1.9 (2) 7.6, 8.4, 7.8, 8.1, 18.4–10.1 (25)
the temperature and velocity magnitudes as functions of y-
10
coordinate at ðx; zÞ ¼ ð0:57; 0:65Þ. This line passes through one of
104 10:25 2.0 (3) 8.3, 8.6, 8.3, 8.6, 22.7–11.3 (29)
the primary branches of the geometry, as shown in Fig. 4b. Consid-
105 10:28 2.1 (10) 8.4, 8.6, 8.7, 8.2, 15.7–9.9 (34)
10:35 2.1 (7) 7.3, 7.4, 7.5, 8.0, 8.4–7.7 (14) ering the rather crude treatment of boundary conditions in the
106
topology optimisation model, it can nevertheless be seen that there
is very good agreement with the COMSOL verification simulation.
parametrised geometry for Gr ¼ 106 after being imported into The major discrepancy is somewhat surprisingly not found at an
COMSOL and the corresponding surface mesh used for the verifica- interface defined by the optimised structure, but rather at the
tion analysis. A full time-dependent simulation carried out for the outer wall (y ¼ 1). Here the finer mesh of the topology optimisa-
Gr ¼ 106 design shows that a steady state exists. Thus, steady state tion model captures the boundary layer better than what is possi-
analyses are performed for the verifications shown. ble with the COMSOL model and its limited resolution capability.

(a) Geometry (b) Surface mesh

Fig. 4. Parametrised geometry and corresponding surface mesh used in the verification analysis for Gr ¼ 106 . The red line shows the line along which the results are
compared: a line from y ¼ 0 to y ¼ 1 at ðx; zÞ ¼ ð0:57; 0:65Þ.
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891 883

0.5
COMSOL
0.45 TopOpt

0.4

0.35

0.3

Temperature
0.25

0.2

0.15

0.1

0.05

0
0 0.05 0.1 0.15 0.2 0.25 0.3
(a) Temperature distribution Position, y-direction
(a) Temperature
0.5
COMSOL
TopOpt 150
0.45

0.4

0.35

0.3
Temperature

Velocity magnitude 100

0.25

0.2

0.15
50
0.1

0.05

0
0 0.2 0.4 0.6 0.8 1
Position, y-direction 0
0 0.05 0.1 0.15 0.2 0.25 0.3
(b) Temperature along line Position, y-direction
(b) Velocity magnitude
300
110
250 100

90
200
Velocity magnitude

80

70
Conductivity

150
60

50
100
40

30
50
20

0 10
0 0.2 0.4 0.6 0.8 1
Position, y-direction 0
0 0.05 0.1 0.15 0.2 0.25 0.3
(c) Velocity magnitude along line Position, y-direction
(c) Conductivity
Fig. 5. Temperature distribution for the verification analysis for Gr ¼ 106 , as well as
temperature and velocity fields as a function of y-coordinate at ðx; zÞ ¼ ð0:57; 0:65Þ.
Fig. 6. Close up of temperature, velocity and conductivity fields as a function of y-
coordinate at ðx; zÞ ¼ ð0:57; 0:65Þ.
884 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891

Fig. 6 shows a close up of the boundary layers around the design thicker heat fins can also be observed in the optimised topologies.
boundaries. Here it can again be seen that the temperature and However, the obtained topologies indicate more than that. If the
velocity profiles differ slightly. The most significant difference is above effect was the only driver for the optimisation, a simple scaled
for the velocity boundary layer. This difference is mainly due to version of the Gr ¼ 103 design (Fig. 2a) with shorter and thicker, but
the Brinkman penalisation used to weakly impose zero velocities mainly circular cross-section, fins would have been obtained for
inside solid parts and no-slip conditions at the interface. The differ- higher Grashof numbers. However, for Gr ¼ 106 (Fig. 2d) one clearly
ences in the temperature level, as well as the objective values in observes more vertically oriented walls than circular fin geometries,
Tables 5 and 7, are due to intermediate material at the interface, indicating that the effect of accelerating the fluid at vertical walls
being upgraded to fully conductive material when thresholded, plays a major role for higher Grashof numbers. This effect is further
as clearly seen from Fig. 6c. The 20% decrease in objective values illustrated in Fig. 7, where surface temperature, flux and local Nus-
for the COMSOL models are mainly attributed to this thresholding
selt number are plotted for the designs obtained for Gr ¼ 103 and
step. Despite these smaller discrepancies, it is concluded that the
topology optimisation model and the associated boundary interpo- Gr ¼ 106 , both evaluated at Gr ¼ 106 . Clearly, the design obtained
lation model is sufficiently accurate to capture the main effects and for Gr ¼ 106 has lower temperature of the heat source and higher
ensure that post-processed designs indeed perform as predicted. fluxes as expected. On the other hand, the design obtained for
If one, despite the above conclusions, would find it pertinent to Gr ¼ 103 is seen to have very cold tips, indicating that its elongated
produce topology optimised structures with sharper edge features fins are inefficient for the higher Grashof number. Also, the local con-
and even better quantitative agreement in verification runs, sev- vection coefficients described by Nusselt numbers is seen to be
eral techniques may be investigated. These techniques include pro- rather constant around and along the fins, whereas the similar val-
jection methods that ensure length scale (to ensure certain ues for the Gr ¼ 106 design are much more non-uniformly dis-
thickness of solid members and fluid channels) [54–56] or inter- tributed, indicating that the fluid flow effects are exploited by the
face capturing schemes, such as X-FEM [38] or discrete simplicial optimised topology to enhance overall heat transfer.
complexes (DSC) [57,58]. It is important to state, that despite inter- Finally, it should be noted that the obtained temperature distri-
face capturing schemes providing more accurate modelling, they bution is far from being constant, which is not unexpected. Consid-
also carry with them many difficulties with regard to the optimisa- ering simple forced convection heat fins, the optimal temperature
tion [59,28,16]. It should be noted that in the authors’ experience, a profile is linearly decaying with the distance from the heat source
small region of intermediate design variables (ensuring a continu- and reaches zero at the tip. For natural convection heat fins, a zero
ous transition from solid to fluid in all material parameters) is ben- temperature difference at the tip is unwanted since this will not
eficial for numerical stability and accuracy and hence beneficial for accelerate the fluid. On the other hand, a constant temperature is
the overall convergence of the optimisation algorithm. not optimal either, since this would imply zero conductive heat
Based on the verification runs in COMSOL, further interpretations flux in the solid parts. Hence, the optimal topology is an intricate
and insight for the obtained results are given in the following. Table 8 compromise between these conflicting goals and obviously
shows the Nusselt and Biot numbers, as well as thermal resistances depends strongly on Biot number. The authors conject that topol-
obtained using COMSOL (these numbers and how they are com- ogy optimisation is an ideal tool to investigate this interplay in
puted are explained in Appendix C). From the table, it can be seen deeper detail. This study will be carried out in future work.
that both conductive and convective resistances are decreased for
increasing Grashof number. The latter effect is expected due to the
6.5. High resolution design
increased flow velocity, whereas the former is due to the more com-
pact geometries obtained for higher Grashof numbers. Their com-
The topology optimisation problem is now investigated with a
bined effect results in increased Biot numbers (ratio between
computational mesh of 320  640  320 elements yielding a total
convection and conduction) and also increased effective Nusselt
of 65; 536; 000 elements and 330; 246; 405 degrees of freedom for
number (average convection coefficient) with increased Grashof
the quarter domain. The design domain now consists of
number. All these tendencies are to be expected and the well-
27; 648; 000 elements and the filter radius is set to 2.5 times the
known fact that increasing Biot number results in shorter and
element size, i.e. in absolute measures half the size of before. The
Table 7
optimisation is run for 1000 design iterations and the computa-
Cross-check objective function values for the verification of designs shown in Fig. 2 tional time was 107 and 108 hours, respectively, for Gr ¼ 103 and
using COMSOL. Bold figures highlight the minimum objective function for a given
Gr ¼ 106 , using 2560 cores. This yields an average of 6.4 and
analysis Grashof number.
6.5 minutes per design iteration, respectively.
Analysis Gr Optimisation Gr Fig. 8 shows the optimised design for Gr ¼ 106 with the fine
103
10 4
10 5
10 6
mesh resolution and small length scale at various steps of the con-
7.00 7.10 7.96 8.57 tinuation strategy. The complexity of the design can be seen to be
103
104 6.54 6.45 6.98 7.51 significantly higher than for the design with a larger length scale,
105 4.90 4.86 4.73 4.87 Fig. 2d. The subfigures are the final iterations of the 1st, 2nd, 3rd
106 3.63 3.65 3.46 3.34 and 5th (final) continuation steps. It can be seen that the complex-
ity of the design decreases during the optimisation process once
the overall topology has been settled (iteration 400 and onwards).
Table 8 This is due to the harder and harder penalisation of intermediate
Biot numbers and thermal resistances for the verification of designs shown in Fig. 2.
design field values, with respect to conductivity, which are present
Gr at the interface between solid and fluid. Therefore, smaller features
are progressively removed as the surface area is more heavily
103 104 105 106
penalised. The reason for going to such high penalisation of the
Rcond 8:76  102 8:50  102 7:21  102 6:69  102 conductivity is to ensure the approximate collocation of the fluid
Rconv 1:92  101 1:73  101 1:17  101 6:69  102 and thermal boundaries, as shown in Section 6.4. However, if one
Bires 0.46 0.49 0.62 1.00
6:72 7:69 11:13 19:59
starts directly with these physically-relevant parameters, particu-
NuL
larly poor local minima have been observed.
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891
(a) Temperature - Gr = 103 design
(c) Normal flux - Gr = 103 design
(e) Local NuL - Gr = 103 design
Fig. 7. Surface temperature, normal flux and local NuL at Gr ¼ 106 for the designs optimised for Gr ¼ 103 and Gr ¼ 106 . Units should be disregarded since the simulations have
been performed using the same dimensionless parameters as presented in the article.
Fig. 9 shows the final optimised designs for Gr ¼ 103 and
Gr ¼ 106 with the fine mesh resolution and small length scale.
The complexities of both designs can be seen to be significantly
higher than the previous, Fig. 2a and d, due to the smaller length
scale. The obtained topologies confirm the results from the coarse
mesh studies in Section 6.3. For low Grashof number the topology
consists of simple circular fins (with occasional fingers), whereas
the topology obtained for the higher Grashof number constitutes
a more complex geometry reflecting the compromise between
compactness, surface area and vertical walls for improved flow
acceleration.
6.6. Change in boundary conditions
An additional problem is considered, where the boundary con-
ditions for the side walls are changed to insulated. These boundary
885
(b) Temperature - Gr = 106 design
(d) Normal flux - Gr = 106 design
(f) Local NuL - Gr = 106 design
conditions were investigated in-depth by Coffin and Maute [38] in
two dimensions, and briefly in three dimensions. Here, a three-
dimensional equivalent of the problem is investigated using a
cubic domain. The full domain is analysed, as it was observed in
[38] that an asymmetric design is beneficial for these boundary
conditions. In order to trigger asymmetric details, the initial design
field is seeded with an angular dependency in the xz-plane with
relation to the x-axis.
The computational mesh is 160  160  160 elements yielding
a total of 4; 096; 000 elements and 20; 866; 405 degrees of freedom
for the full domain. The design domain consists of 1; 728; 000 ele-
ments and the filter radius is set to 2.5 times the element size. The
optimisation is run for 500 design iterations and the computational
time was 10 hours 37 minutes and 10 hours 53 minutes, respec-
tively, for Gr ¼ 103 and Gr ¼ 104 , using 640 cores. This yields an
average of approximately 1.3 minutes per design iteration.
886 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891
(a) Iteration 200
(c) Iteration 600
6
Fig. 8. Optimised designs for Gr ¼ 10 at a mesh resolution of 320  640  320. Please note that the freely hanging material in (a) is due to only elements below the threshold,
c ¼ 0:05, are shown – the design is connected by intermediate design field values.
Fig. 10 shows the final optimised designs for Gr ¼ 103 and
Gr ¼ 104 . It is interesting to see that the asymmetry of the
optimised designs appears to depend on the Grashof number.
This was also observed in two dimensions by Coffin and Maute
[38]. In order to compare the asymmetric designs to equivalent
symmetric designs, the problem is also analysed by enforcing
quarter symmetry. Fig. 11 shows the final optimised quarter
symmetric designs for Gr ¼ 103 and Gr ¼ 104 . By comparing
Figs. 10 and 11, it can be seen that for Gr ¼ 103 , the design
remains close to symmetric, but as convection-dominance is
increased, Gr ¼ 104 , the design becomes asymmetric to a large
degree.
Table 9 compares the objective function values for enforced
quarter symmetry and the free design. It can be seen that the
asymmetric designs perform better for both Grashof numbers,
but that the difference is most significant for Gr ¼ 104 .
The designs observed for Gr ¼ 103 is similar in appearance as
the three-dimensional design in [38] for very similar parameter
values. The design clearly shows that the goal is now to conduct the
heat to the very top of the domain for the diffusion-dominant
(b) Iteration 400
(d) Iteration 1000
problem, but this quickly changes for increasing convection-
dominance where the possibility of an asymmetric flow is taken
advantage of.
The problem has only been investigated for Gr ¼ 103 and
Gr ¼ 104 , as convergence issues (non-linear state solver) were
encountered for higher Grashof numbers when treating the full
domain. However, when enforcing quarter symmetry (in both the
design and state solutions), higher Grashof numbers (Gr ¼ 105
and Gr ¼ 106 ) could be optimised without problems. This is likely
due to the fact that the flow becomes unstable for higher Grashof
numbers when allowing for asymmetric designs. In order to opti-
mise for these conditions, one has to use a transient solver and
optimisation routine, e.g. [26,38].
It can be seen that for this particular choice of boundary condi-
tions, an asymmetric design performs better than a quarter sym-
metric design. This is not the case when the side walls are kept
at the same cold temperature as the top, Fig. 2. The adiabatic walls
allow for the flow parallel to the wall to move in both the upward
and downward directions. However, when they are cold, as for the
initial problem, the flow is essentially forced to descend at the side
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891 887

(a) Gr = 103 (b) Gr = 106

Fig. 9. Optimised designs for Gr ¼ 103 and Gr ¼ 106 at a mesh resolution of 320  640  320. The outline of the outer walls are shown in black.

Fig. 10. Optimised designs for Gr ¼ 103 and Gr ¼ 104 for cavity with insulated sides.
888 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891
Fig. 11. Optimised quarter symmetric designs for Gr ¼ 103 and Gr ¼ 104 for cavity with insulated sides.
Table 9
Objective function values for the designs in cavity with insulated sides.
Symmetry Gr
103
Enforced 22.64
Not enforced 22.25
walls. The adiabatic walls allows for the fluid to ascend on some
sides and the flow to collect into a single main convection roll, in
turn minimising viscous losses. With cold side walls, the flow is
pretty much hard set to four main convection rolls with the solid
member in the middle. It is interesting to note, that contrary to
observations in two dimensions [38], the maximum velocity is sig-
nificantly lower (37.5) for the asymmetric design than for the quar-
ter symmetric design (50.7). Based on the observations from the
steady state solver regarding stability, it is postulated that the
extra constrained conditions, may also provide a stabilising effect
and move the critical Grashof number upwards, allowing for
steady-state solutions at higher Grashof numbers.
7. Discussion and conclusion
In this article, topology optimisation has been applied to the
design of three-dimensional heat sinks using a fully coupled non-
linear thermofluidic model. In contrast to previous works within
topology optimisation, that considered simplified convection mod-
els, the presented methodology is able to recover interesting phys-
ical effects and insights, and avoids problems with the formation of
non-physical internal cavities, length-scale effects and artificial
convection assumptions. The implementation of the code in a
PETSc framework suitable for large scale parallel computations
allows for running examples with more than 300 million degrees
of freedom and almost 30 million design variables on regular grids.
The obtained topology optimisation results are verified with the
commercial finite element code COMSOL and show consistent
results. Field distributions from the hexahedral mesh used in the
topology optimisation are in good agreement with results obtained
with a smoothed geometry and body-fitted mesh in COMSOL. Fur-
thermore, optimisation results are consistent, i.e. topologies opti-
mised for a certain Grashof number are also the best for that
Grashof number when modelled in Comsol.
The examples considered to verify the presented framework are
primarily academic in nature. Nevertheless, some interesting
insight is obtained, showing that optimised structures go from
exhibiting simple branches that conduct heat towards the cold
104 outer boundaries for diffusion-dominated problems, towards com-
14.27 plex and compact multi-branched structures that maximise the
13.17 convection heat transfer for higher Grashof numbers. The observa-
tion that lower rates of natural convection produce fewer longer
fins, while higher rates of natural convection produce a greater
number of short fins/branches are consistent with prior optimisa-
tion literature. However, the optimisation also suggests structures
that are good at accelerating fluid, in turn ensuring enhanced over-
all heat transfer, and it should be stressed that here the designs
appear spontaneously, without any predefined knowledge or
restriction on the design.
Current and future work includes applications to real life prob-
lems (see preliminary work in [39]), irregular meshes, multiple ori-
entations, as well as the extension to transient and turbulent
problems. Investigations into enhanced accuracy of the boundary
layer for highly convection-dominated problems is also of primary
interest.
Acknowledgements
This work was funded by Villum Fonden through the NextTop
project, as well as by Innovation Fund Denmark through the
HyperCool project. The first author was also partially funded by
the EU FP7-MC-IAPP programme LaScISO.
The anonymous reviewers are thanked for their insightful sug-
gestions for improving the article.
Appendix A. Stabilisation parameters
The stabilised weak form of the governing equations are given
in [18]. The governing equations are stabilised using the
pressure-stabilising Petrov–Galerkin (PSPG) [60,61] and
streamline-upwind Petrov–Galerkin (SUPG) methods [62], for
more information please see [18].
The stabilisation parameters for SUPG and PSPG are defined as
one and the same using the following approximate min-function:

12
1 1 1
sSU ¼ sPS ¼ s ¼ þ þ ðA:1Þ
s1
2
s2
2
s23
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891 889

The limit factors are given by: Table 10

Computational and actual Gr-numbers for the designs presented in this article.
4he
s1 ¼ ðA:2aÞ Computational Gr Mesh Figure Actual Gr
jjue jj2
10 3 160  320  160 2a 1:69  103
2
he 160  320  160
s2 ¼ ðA:2bÞ 104 2b 1:55  104
4Pr 105 160  320  160 2c 1:16  105
1 106 160  320  160 2d 8:60  105
s3 ¼ ðA:2cÞ
ae 103 320  640  320 9a 1:42  103
106 320  640  320 9b 7:16  105
where he is a characteristic element size (for cubes the element 103 160  160  160 10a 4:49  103
edge length) and ue is the element vector of velocity degrees of 104 160  160  160 10b 2:68  104
freedom. The first limit factor, s1 , has been simplified based on 103 80  160  80 11a 4:57  103
evaluation at element centroids under the assumption of a single 104 80  160  80 11b 2:89  104
Gauss-point, yielding a constant stabilisation factor within each
element.
In order to define a consistent Jacobian matrix, and thus a Working with finite precision arithmetic, this of course cannot be
consistent adjoint problem, the derivatives of the stabilisation guaranteed.
factors are needed with respect to the velocity field. This can be The actual Grashof numbers for the designs shown in this arti-
found to be: cle are listed in Table 10. The values for the finer meshes are lower

2
2 !1 due to the lower thermal compliance, equivalent to the tempera-
@s s1 s1  1 ture of the heat source, allowed by the smaller design length scale.
¼ s 1 þ þ ue T ue uTe ðA:3Þ
@ue s2 s3
Appendix C. Calculation of thermal resistances and Nusselt
Furthermore, to define consistent design sensitivities, the number
derivatives of the stabilisation factors with respect to the design
field is needed. This can be found to be: Asterisks denote dimensional quantities related to the un-

2
2 !1 asterisked non-dimensional equivalents. It is assumed, without
@s s s3 s3 @ ae loss of generality, that the free flow temperature and the reference
¼ þ þ1 ðA:4Þ
@ ce s3 s1 s2 @ ce temperature, T 0 , are the same.

Including the derivatives in the definition of consistent adjoint

C.1. Thermal resistance analysis
sensitivities has been observed to provide a one to two order of
magnitude improvement in accuracy of sensitivities with respect
It is assumed that the heat loss from the sides of the heat source
to a finite difference approximation. Significant differences have
is neglectable, to simplify the analysis. Using the verification calcu-
not been observed in optimisation behaviour or in final designs.
lations, this has been confirmed. All power from the heat source is
However, this cannot be guaranteed in general and it is therefore
thus assumed to pass through the heat sink surface. For an intro-
best to ensure consistency, as also highlighted by the similar issue
duction to thermal resistance analysis, please see e.g. [64].
of frozen turbulence [63].
The conductive resistance is defined as:
Similar definitions and derivations are carried out for the ther-
mal SUPG stabilisation. T b  T s
Rcond ¼ ðC:1Þ
Q
Appendix B. Computational versus actual Grashof number where T b is the temperature of the base (heat source) and T s is the
temperature of the heat sink surface. Furthermore, Q  is the total
The Grashof numbers used above are all based on an a priori power of the heat source:
defined reference temperature difference. However, the actual Z

temperature difference observed between the heat source and Q ¼ q dV ¼ q x ðC:2Þ
the walls are not known a priori due to the fact that the problem x
only has a single known temperature (Dirichlet boundary condi- where x is the volume of the heat source.
tion) and a volumetric heat source. This is why the maximum tem- Eq. C.1 is non-dimensionalised using the same quantities as the
perature observed for the designs, see Fig. 2, is not equal to 1. governing equations [18], yielding:
Therefore, after the optimisation, one can define an a posteriori
1
Grashof number based on the actual computed temperature differ- Rcond ¼ Rcond ðC:3Þ
ence. The a priori version can be termed the computational Grashof
kf L
number, the one that goes into the dimensionless governing equa- where the non-dimensional conductive resistance is defined as:
tions; and the a posteriori version can be called the actual, or opti-
Tb  Ts
mised, Grashof number for the given optimised design. The actual Rcond ¼ ðC:4Þ
Grashof number can be useful for experimental studies, as well as s0 x
for future comparisons. An equivalent non-dimensionalisation is performed for the con-
The temperature scale used in the calculation of the Grashof vective resistance:
number is arbitrary for computational purposes, as it merely
T s  T 0 1
scales the non-dimensional temperature field. Thus, if a smaller Rconv ¼ ¼ Rconv ðC:5Þ
temperature scale is used, the non-dimensional temperature Q kf L
increases proportionally; and if a larger temperature scale is used, where the non-dimensional convective resistance is defined as:
the non-dimensional temperature decreases proportionally.
Assuming full precision arithmetic, the final designs, obtained Ts
Rconv ¼ ðC:6Þ
for two arbitrarily chosen temperature scales, should be the same. s0 x
890 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891

According to e.g. [64], the Biot number can be defined as the [5] D. Jang, S.-H. Yu, K.-S. Lee, Multidisciplinary optimization of a pin-fin radial
heat sink for led lighting applications, Int. J. Heat Mass Transfer 55 (4) (2012)
ratio of the conductive resistance to convective resistance. This
515–521, http://dx.doi.org/10.1016/j.ijheatmasstransfer.2011.11.016.
gives: [6] A. Bejan, Constructal-theory network of conducting paths for cooling a heat
generating volume, Int. J. Heat Mass Transfer 40 (4) (1997) 799–816, http://dx.
Rcond Rcond T b  T s doi.org/10.1016/0017-9310(96)00175-5.
Bires ¼ ¼ ¼ ðC:7Þ
Rconv Rconv Ts [7] M.P. Bendsøe, O. Sigmund, Topology Optimization: Theory, Methods and
Applications, Springer, 2003. ISBN: 3-540-42992-1.
[8] O. Sigmund, K. Maute, Topology optimization approaches, Struct. Multidiscip.
Looking at this expression clearly shows that Bires represents a mea- Optim. 48 (6) (2013) 1031–1055, http://dx.doi.org/10.1007/s00158-013-0978-6.
sure of the isothermality of the heat sink geometry. [9] J.D. Deaton, R.V. Grandhi, A survey of structural and multidisciplinary
In order to use the above expressions for the complex three- continuum topology optimization: post 2000, Struct. Multidiscip. Optim. 49
(1) (2013) 1–38, http://dx.doi.org/10.1007/s00158-013-0956-z.
dimensional heat sink geometries presented, T b is defined to be
[10] L. Yin, G. Ananthasuresh, A novel topology design scheme for the multi-physics
the volume-averaged temperature of the base (heat source): problems of electro-thermally actuated compliant micromechanisms, Sensors
Z Actuators 97–98 (2002) 599–609, http://dx.doi.org/10.1016/S0924-4247(01)
1 00853-6.
Tb ¼ TdV ðC:8Þ
x x
[11] T. Bruns, Topology optimization of convection-dominated, steady-state heat
transfer problems, Int. J. Heat Mass Transfer 50 (15–16) (2007) 2859–2873,
and T s to be the surface-averaged temperature of the surface: http://dx.doi.org/10.1016/j.ijheatmasstransfer.2007.01.039.
[12] A. Iga, S. Nishiwaki, K. Izui, M. Yoshimura, Topology optimization for thermal
Z conductors considering design-dependent effects, including heat conduction
1
Ts ¼ TdS ðC:9Þ and convection, Int. J. Heat Mass Transfer 52 (11–12) (2009) 2721–2732,
Ahs Chs http://dx.doi.org/10.1016/j.ijheatmasstransfer.2008.12.013.
[13] J. Alexandersen, Topology optimisation for convection problems, Technical
where Chs is the heat sink surface and Ahs the area thereof. University of Denmark, 2011. B.Eng. thesis, <http://orbit.dtu.dk/en/>.
[14] E. Dede, S.N. Joshi, F. Zhou, Topology optimization, additive layer
manufacturing, and experimental testing of an air-cooled heat sink, ASME
C.2. Nusselt number Journal of Mechanical Design 137 (11). doi:10.1115/1.4030989.
[15] M. Zhou, J. Alexandersen, O. Sigmund, C.B.W. Pedersen, Industrial application
of topology optimization for combined conductive and convective heat
Newton’s law of cooling gives the relation between the local transfer problems, Struct. Multidiscip. Optim. (2016) 1–16, http://dx.doi.org/
normal flux, q00;
n , and the convection heat transfer coefficient, h: 10.1007/s00158-016-1433-2. online first.
[16] P. Coffin, K. Maute, Level set topology optimization of cooling and heating

q00;
n ¼ hðT  T 0 Þ ðC:10Þ devices using a simplified convection model, Struct. Multidiscip. Optim. (2015)
1–19, http://dx.doi.org/10.1007/s00158-015-1343-8.
From this, one can isolate the convection coefficient and define an [17] J. Alexandersen, Topology optimisation for axisymmetric convection problems,
Technical University of Denmark, 2011. Tech. rep., <http://orbit.dtu.dk/en/>.
effective heat transfer coefficient, h, as the surface-averaged [18] J. Alexandersen, N. Aage, C.S. Andreasen, O. Sigmund, Topology optimisation of
quantity: natural convection problems, Int. J. Numer. Meth. Fluids 76 (10) (2014) 699–
Z 721, http://dx.doi.org/10.1002/fld.3954.
1 qn00;  [19] T. Borrvall, J. Petersson, Topology optimization of fluids in Stokes flow, Int. J.
h¼ dS ðC:11Þ
Ahs
Numer. Meth. Fluids 41 (1) (2003) 77–107, http://dx.doi.org/10.1002/
Chs T  T0 fld.426.
[20] A. Gersborg-Hansen, O. Sigmund, R. Haber, Topology optimization of channel
Eq. C.11 is non-dimensionalised yielding: flow problems, Struct. Multidiscip. Optim. 30 (3) (2005) 181–192, http://dx.
Z doi.org/10.1007/s00158-004-0508-7.
kf 1 q00n [21] C.S. Andreasen, A.R. Gersborg, O. Sigmund, Topology optimization of
h¼  dS ðC:12Þ
L Ahs Chs T microfluidic mixers, Int. J. Numer. Meth. Fluids 61 (5) (2009) 498–513,
http://dx.doi.org/10.1002/fld.1964.
[22] D. Makhija, K. Maute, Level set topology optimization of scalar transport
From this, the effective Nusselt number, based on the reference
problems, Struct. Multidiscip. Optim. 51 (2) (2015) 267–285, http://dx.doi.org/
length of the problem, can be defined as: 10.1007/s00158-014-1142-7.
Z [23] F. Okkels, H. Bruus, Scaling behavior of optimally structured catalytic
hL 1 q00n microfluidic reactors, Phys. Rev. E 75 (1) (2007) 016301, http://dx.doi.org/
NuL ¼ ¼ dS ðC:13Þ 10.1103/PhysRevE.75.016301.
kf Ahs Chs T [24] S. Kreissl, G. Pingen, K. Maute, Topology optimization for unsteady flow, Int. J.
Numer. Meth. Eng. 87 (13) (2011) 1229–1253, http://dx.doi.org/10.1002/
By looking at the physical arguments and the origin of the expres- nme.3151.
sion, it can be seen that the effective Nusselt number represents a [25] Y. Deng, Z. Liu, P. Zhang, Y. Liu, Y. Wu, Topology optimization of unsteady
incompressible Navier–Stokes flows, J. Comput. Phys. 230 (17) (2011) 6688–
non-dimensional effective heat transfer coefficient.
6708, http://dx.doi.org/10.1016/j.jcp.2011.05.004.
A local Nusselt number equivalent to the local convection coef- [26] S. Nørgaard, O. Sigmund, B. Lazarov, Topology optimization of unsteady flow
ficient, h, is simply defined as: problems using the lattice boltzmann method, J. Comput. Phys. 307 (2016)
291–307, http://dx.doi.org/10.1016/j.jcp.2015.12.023.
hL q00n [27] G.H. Yoon, Topology optimization for stationary fluid-structure interaction
NuL ¼ ¼ ðC:14Þ problems using a new monolithic formulation, Int. J. Numer. Meth. Eng. 82 (5)
kf T (2010) 591–616, http://dx.doi.org/10.1002/nme.2777.
[28] N. Jenkins, K. Maute, Level set topology optimization of stationary fluid–
structure interaction problems, Struct. Multidiscip. Optim. 52 (1) (2015) 179–
References 195, http://dx.doi.org/10.1007/s00158-015-1229-9.
[29] E.A. Kontoleontos, E.M. Papoutsis-Kiachagias, A.S. Zymaris, D.I. Papadimitriou,
K.C. Giannakoglou, Adjoint-based constrained topology optimization for
[1] A.T. Morrison, Optimization of heat sink fin geometries for heat sinks in natural
viscous flows, including heat transfer, Eng. Optim. 45 (8) (2013) 941–961,
convection, in: InterSociety Conference on Thermal Phenomena in Electronic
http://dx.doi.org/10.1080/0305215X.2012.717074.
Systems, I-THERM III, 1992, pp. 145–148, http://dx.doi.org/10.1109/
[30] E. Dede, Multiphysics topology optimization of heat transfer and fluid flow
ITHERM.1992.187753.
systems, in: Proceedings of the COMSOL Conference 2009 Boston, 2009.
[2] G. Ledezma, A. Bejan, Optimal geometric arrangement of staggered vertical
[31] G.H. Yoon, Topological design of heat dissipating structure with forced
plates in natural convection, J. Heat Transfer 119 (4) (1997) 700–708, http://
convective heat transfer, J. Mech. Sci. Technol. 24 (6) (2010) 1225–1233,
dx.doi.org/10.1115/1.2824174.
http://dx.doi.org/10.1007/s12206-010-0328-1.
[3] M. Iyengar, A.B. Cohen, Least-material optimization of vertical pin-fin, plate-
[32] K. Lee, Topology optimization of convective cooling system designs, University
fin, and triangular-fin heat sinks in natural convective heat transfer, in: The
of Michigan, 2012. Ph.D. thesis, <http://deepblue.lib.umich.edu>.
Sixth Intersociety Conference on Thermal and Thermomechanical Phenomena
[33] A.A. Koga, E.C.C. Lopes, H.F.V. Nova, C.R. de Lima, E.C.N. Silva, Development
in Electronic Systems, 1998. ITHERM’98, IEEE, 1998, pp. 295–302.
of heat sink device by using topology optimization, Int. J. Heat Mass
[4] R. Bahadur, A. Bar-Cohen, Thermal design and optimization of natural
Transfer 64 (2013) 759–772, http://dx.doi.org/10.1016/j.ijheatmasstransfer.
convection polymer pin fin heat sinks, IEEE Trans. Compon. Packag. Technol.
2013.05.007.
28 (2) (2005) 238–246, http://dx.doi.org/10.1109/TCAPT.2005.848498.
 J. Alexandersen et al. / International Journal of Heat and Mass Transfer 100 (2016) 876–891
[34] G. Marck, M. Nemer, J.-L. Harion, Topology optimization of heat and mass
transfer problems: Laminar flow, Numer. Heat Transfer, Part B: Fundam. 63 (6)
(2013) 508–539, http://dx.doi.org/10.1080/10407790.2013.772001.
[35] J.H.K. Haertel, K. Engelbrecht, B.S. Lazarov, O. Sigmund, Topology optimization
of thermal heat sinks, in: Proceedings of COMSOL conference 2015, 2015.
[36] K. Yaji, T. Yamada, M. Yoshino, T. Matsumoto, K. Izui, S. Nishiwaki, Topology
optimization in thermal-fluid flow using the lattice Boltzmann method,
J. Comput. Phys. 307 (2016) 355–377, http://dx.doi.org/10.1016/j.jcp.2015.
12.008.
[37] L. Laniewski-Wollk, J. Rokicki, Adjoint lattice Boltzmann for topology
optimization on multi-GPU architecture, Comput. Math. Appl. 71 (3) (2016)
833–848, http://dx.doi.org/10.1016/j.camwa.2015.12.043.
[38] P. Coffin, K. Maute, A level-set method for steady-state and transient natural
convection problems, Struct. Multidiscip. Optim. (2015) 1–21, http://dx.doi.
org/10.1007/s00158-015-1377-y.
[39] J. Alexandersen, O. Sigmund, N. Aage, Topology optimisation of passive coolers
for light-emitting diode lamps, in: 11th World Congress on Structural and
Multidisciplinary Optimization, 2015. <www.aeromech.usyd.edu.au/
WCSMO2015/papers/1264paper.pdf>.
[40] E. Andreassen, A. Clausen, M. Schevenels, B.S. Lazarov, O. Sigmund, Efficient
topology optimization in MATLAB using 88 lines of code, Struct. Multidiscip.
Optim. 43 (2011) 1–16, http://dx.doi.org/10.1007/s00158-010-0594-7.
[41] O. Amir, N. Aage, B.S. Lazarov, On multigrid-CG for efficient topology
optimization, Struct. Multidiscip. Optim. (2013) 1–15, http://dx.doi.org/
10.1007/s00158-013-1015-5.
[42] J. Alexandersen, B.S. Lazarov, Topology optimisation of manufacturable
microstructural details without length scale separation using a spectral
coarse basis preconditioner, Comput. Methods Appl. Mech. Eng. 290 (2015)
156–182, http://dx.doi.org/10.1016/j.cma.2015.02.028.
[43] T.H. Nguyen, G.H. Paulino, J. Song, C.H. Le, Improving multiresolution topology
optimization via multiple discretizations, Int. J. Numer. Meth. Eng. 92 (6)
(2012) 507–530, http://dx.doi.org/10.1002/nme.4344.
[44] T. Borrvall, J. Petersson, Large-scale topology optimization in 3D using parallel
computing, Comput. Methods Appl. Mech. Eng. 190 (2001) 6201–6229, http://
dx.doi.org/10.1016/S0045-7825(01)00216-X.
[45] A. Evgrafov, C.J. Rupp, K. Maute, M.L. Dunn, Large-scale parallel topology
optimization using a dual-primal substructuring solver, Struct. Multidiscip.
Optim. 36 (2008) 329–345, http://dx.doi.org/10.1007/s00158-007-0190-7.
[46] N. Aage, B. Lazarov, Parallel framework for topology optimization using the
method of moving asymptotes, Struct. Multidiscip. Optim. 47 (4) (2013) 493–
505, http://dx.doi.org/10.1007/s00158-012-0869-2.
[47] N. Aage, E. Andreassen, B.S. Lazarov, Topology optimization using PETSc: an
easy-to-use, fully parallel, open source topology optimization framework,
Struct. Multidiscip. Optim. 51 (3) (2014) 565–572, http://dx.doi.org/10.1007/
s00158-014-1157-0.
[48] S. Balay, S. Abhyankar, M.F. Adams, J. Brown, P. Brune, K. Buschelman, V.
Eijkhout, W.D. Gropp, D. Kaushik, M.G. Knepley, L.C. McInnes, K. Rupp, B.F.
Smith, H. Zhang, PETSc Users Manual, Tech. Rep. ANL-95/11 – Revision 3.5,
Argonne National Laboratory (2014). <http://www.mcs.anl.gov/petsc>.
[49] B.S. Lazarov, O. Sigmund, Filters in topology optimization based on Helmholtz-
type differential equations, Int. J. Numer. Meth. Eng. 86 (2011) 765–781,
http://dx.doi.org/10.1002/nme.3072.
891
[50] K. Svanberg, The method of moving asymptotes – a new method for structural
optimization, Int. J. Numer. Meth. Eng. 24 (2) (1987) 359–373, http://dx.doi.
org/10.1002/nme.1620240207.
[51] M. Stolpe, K. Svanberg, On the trajectories of penalization methods for
topology optimization, Struct. Multidiscip. Optim. 21 (2001) 128–139, http://
dx.doi.org/10.1007/s001580050177.
[52] S. Rojas-Labanda, M. Stolpe, Benchmarking optimization solvers for structural
topology optimization, Struct. Multidiscip. Optim. 52 (3) (2015) 527–547,
http://dx.doi.org/10.1007/s00158-015-1250-z.
[53] P.S. Vassilevski, Multilevel Block Factorization Preconditioners, Springer
(2007), http://dx.doi.org/10.1007/978-0-387-71564-3.
[54] F. Wang, B.S. Lazarov, O. Sigmund, On projection methods, convergence and
robust formulations in topology optimization, Struct. Multidiscip. Optim. 43
(6) (2011) 767–784, http://dx.doi.org/10.1007/s00158-010-0602-y.
[55] M. Zhou, B.S. Lazarov, F. Wang, O. Sigmund, Minimum length scale in topology
optimization by geometric constraints, Comput. Methods Appl. Mech. Eng. 293
(2015) 266–282, http://dx.doi.org/10.1016/j.cma.2015.05.003.
[56] B.S. Lazarov, F. Wang, O. Sigmund, Length scale and manufacturability in
density-based topology optimization, Arch. Appl. Mech. (2016) 1–30, http://
dx.doi.org/10.1007/s00419-015-1106-4.
[57] A.N. Christiansen, M. Nobel-Jrgensen, N. Aage, O. Sigmund, J.A. Brentzen,
Topology optimization using an explicit interface representation, Struct.
Multidiscip. Optim. 49 (3) (2013) 387–399, http://dx.doi.org/10.1007/
s00158-013-0983-9.
[58] A.N. Christiansen, J.A. Bærentzen, M. Nobel-Jørgensen, N. Aage, O. Sigmund,
Combined shape and topology optimization of 3D structures, Comput.
Graphics 46 (2015) 25–35, http://dx.doi.org/10.1016/j.cag.2014.09.021.
shape Modeling International 2014.
[59] D. Makhija, K. Maute, Numerical instabilities in level set topology optimization
with the extended finite element method, Struct. Multidiscip. Optim. 49 (2)
(2013) 185–197, http://dx.doi.org/10.1007/s00158-013-0982-x.
[60] T.J. Hughes, L.P. France, M. Balestra, A new finite element formulation for
computational fluid dynamics V - circumventing the Babuska–Brezzi
condition: a stable Petrov–Galerkin formulation of the Stokes problem
accommodating equal-order interpolations, Comput. Methods Appl.
Mech. Eng. 59 (1) (1986) 85–99, http://dx.doi.org/10.1016/0045-7825(86)
90025-3.
[61] T.E. Tezduyar, S. Mittal, S. Ray, R. Shih, Incompressible flow computations with
stabilized bilinear and linear equal-order-interpolation velocity–pressure
elements, Comput. Methods Appl. Mech. Eng. 95 (2) (1992) 221–242, http://
dx.doi.org/10.1016/0045-7825(92)90141-6.
[62] A.N. Brooks, T.J. Hughes, Streamline upwind/Petrov–Galerkin formulations for
convection dominated flows with particular emphasis on the incompressible
Navier–Stokes equations, Comput. Methods Appl. Mech. Eng. 32 (1–3) (1982)
199–259, http://dx.doi.org/10.1016/0045-7825(82)90071-8.
[63] A.S. Zymaris, D.I. Papadimitriou, K.C. Giannakoglou, C. Othmer, Continuous
adjoint approach to the Spalart-Allmaras turbulence model for incompressible
flows, Comput. Fluids 38 (2009) 1528–1538, http://dx.doi.org/10.1016/
j.compfluid.2008.12.006.
[64] F.P. Incropera, D.P. DeWitt, T.L. Bergman, A.S. Lavine, Introduction to Heat
Transfer, fifth ed., Wiley, 2007. iSBN: 978-0-471-45727-5.